RedCaps is a large-scale image-text pair dataset with 1.2 million images and texts from Reddit. These images and texts describe a variety of objects and scenes. The data was collected from a set of human-curated subreddits that provided coarse image labels and allowed for guiding the assembly of the dataset without labeling individual instances. The team at the University of Michigan released the dataset.